[Paper Review] GraphCast: Learning skillful medium-range global weather forecasting

Rémi Lam, Álvaro Sánchez‐González+16 more|arXiv (Cornell University)|Dec 24, 2022

Advanced Graph Neural Networks169 citations

TL;DR

GraphCast is a graph neural network–based meteorological model trained on ERA5 reanalysis data to predict hundreds of weather variables up to 10 days at 0.25° resolution, outperforming ECMWF HRES on the majority of targets. It runs fast and supports severe-event prediction.

ABSTRACT

Global medium-range weather forecasting is critical to decision-making across many social and economic domains. Traditional numerical weather prediction uses increased compute resources to improve forecast accuracy, but cannot directly use historical weather data to improve the underlying model. We introduce a machine learning-based method called "GraphCast", which can be trained directly from reanalysis data. It predicts hundreds of weather variables, over 10 days at 0.25 degree resolution globally, in under one minute. We show that GraphCast significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets, and its forecasts support better severe event prediction, including tropical cyclones, atmospheric rivers, and extreme temperatures. GraphCast is a key advance in accurate and efficient weather forecasting, and helps realize the promise of machine learning for modeling complex dynamical systems.

Motivation & Objective

• Demonstrate that a data-driven MLWP model can be trained directly from reanalysis data to forecast global weather variables up to 10 days ahead.
• Show that GraphCast can outperform the operational deterministic system HRES on a large set of verification targets.
• Evaluate GraphCast's usefulness for severe-event prediction (cyclones, atmospheric rivers, extreme temperatures).
• Assess the impact of training data recency on forecast skill and discuss limitations and future directions.

Proposed method

• Encode–process–decode GraphCast architecture using a multi-mesh graph derived from refined icosahedral meshes.
• Autoregressive forecasting by predicting state residuals over 6-hour steps.
• Train on 39 years (1979–2017) of ERA5 reanalysis with MSE loss weighted by vertical level, progressively increasing autoregressive steps from 1 to 12 steps.
• Compare GraphCast to HRES using RMSE and anomaly correlation coefficient (ACC) across 1380 targets for 10-day forecasts at 0.25° resolution.
• Use ground-truth grounding via ERA5 as input/output, and construct an HRES-fc0 baseline to ensure fair comparison with similar data-assimilation windows.
• Operate GraphCast with 36.7 million parameters on Google Cloud TPUs and rely on a 0.25° × 0.25° grid with 37 pressure levels.

Experimental results

Research questions

• RQ1 Can a graph neural network trained on ERA5 reanalysis data learn to predict weather states across many variables up to 10 days ahead?
• RQ2 How does GraphCast perform relative to ECMWF’s HRES deterministic model on a wide set of verification targets across lead times?
• RQ3 Does GraphCast provide improvements in severe-event forecasting (cyclones, atmospheric rivers, extreme temperatures) even though not directly trained for them?
• RQ4 What is the impact of using recent training data on GraphCast’s forecast skill?
• RQ5 What are the limitations of a deterministic ML-based forecast compared to ensemble methods?

Key findings

• GraphCast outperforms HRES on 90.3% of 1380 targets and 89.9% significantly, across 10-day forecasts at 0.25° resolution.
• For the z500 headline field, GraphCast shows 7–14% RMSE skill improvement and higher ACC than HRES.
• GraphCast maintains superior performance across most variables and levels; removing the highest/lowest 50 hPa levels raises dominance to 96.9% (excluding 50 hPa) and 99.7% (excluding 50 and 100 hPa).
• GraphCast’s cyclone-tracks are more accurate than HRES for lead times 18 hours to 4.75 days.
• GraphCast improves atmospheric-river related IVT predictions compared with HRES, particularly at longer lead times (roughly 10–25% RMSE improvement).
• Against extreme-heat/cold event predictions, GraphCast shows better precision-recall at 5–10 day leads, though HRES can be better at 12 hours.